Methods Of Feature Dimensionality Reduction Based On Semi-supervised Learning

Feature dimension reduction is a important direction in machine learning field, widely used in computer vision, data mining, and so on. Feature reduction aims at removing a lot of redundant information that is present in the original data and to find out the effective intrinsic characteristics in the complex data. In practice, large number of unlabeled samples are readily available, while, there are a limited number of labeled samples. For this case the traditional supervised and unsupervised feature dimension reduction methods can not get the good results. This thesis focuses on the feature dimension reduction methods based on semi-supervised learning, to effectively utlize the information from unlabeled samples.Firstly, aiming at the limitation of supervised intra-class graphs and inter-class graphs in TCA, which cannot take full advantage of the class label information issues, we introduce the discriminant analysis and propose a Transductive Local Fisher Discriminant Analysis(TLFDA) that combines the TCA and Local Fisher Discriminant Analysis together. This method makes class label information of samplesmore fully utilized by calculating the local within-class scatter matrix and the local between-class scatter matrix, and at the same time keeps that the local structure information of samples can be maintained the characteristics of TCA method. Experimental results show that the Transductive Local Fisher Discriminant Analysis perform better than traditional dimension reduction method based on semi-supervised learning.Secondly, inview of the limitation of SELF that the local structure information of only a small amount of labeled samples can be maintained. We proposed Transductive discriminant analysis(TLDA) that combines linear discriminant analysis and improved LPP together. This method caneffectively taken into account the large number of the local structure information of unlabeled samples and the global information of labeled samples. Experimental results show that Transductive discriminant analysis perform better than traditional dimension reduction method based on semi-supervised learning.Finally, aiming at inview of the problem that the traditional Graph don't have label information, we propose a new supervised projection method to build Graph. Then k-Nearest Neighbors Graph and supervised indirect graph are projected onto the new feature space to obtain supervised projection graph. Further, we build an indirect inter-class label graph by using label information to make a different class samples away from each other. We put forward a new Transductive Locality Preserving Projection which is based on supervised projection graph and indirect inter-class label graph. This method maintains the local structure information of the entire sample. The experiments verify that the new semi-supervised dimension reduction method based on supervised projection graph performs better than traditional dimension reduction method based on semi-supervised learning.